AI Agent Workflow Collection

When the persona needs to watch a long-running local subprocess for completion or stream events from it without burning conversation turns on busy-waiting, reach for the Monitor tool. Monitor is event-driven (wakes the agent when the watched condition fires), strictly more token-efficient than a /loop poll, and the right primitive for 'wait until X happens' on local artifacts (file mtime, ps liveness, log-line match, port-open). Use when: waiting for a background-agent build to finish, watching a log file for an error pattern, blocking on a port becoming available, or any 'fire when condition holds' pattern against local state. Composes with run-in-background-bash (which detaches the work; Monitor watches it) and the dead-agent-detection memory rule (which obligates probe-the-artifact over poller-cadence inference).

Use when the user explicitly asks to capture a pattern as a skill ('remember this', 'make this a skill', 'let's codify this', 'capture this as a skill'), OR when the same multi-step procedure is invoked 3+ times within a session, OR when the same shape of question is asked 3+ times and the answer text stabilizes (ask-and-answer pattern). The persona drafts a workspace-local SKILL.md to a .scratch/ path, surfaces a one-line ratification question through PM (Y/N/R), and on Y writes the SKILL to /.claude/skills/ /SKILL.md so Claude Code's native discovery auto-loads it next turn. Universal — any loam user, dev or non-dev. Gated on the workspace-bootstrap manifest's enableautoskillcapture flag (default false; user opts in when ready).